Robust Automatic 3D Brain Extraction on T1 Weighted Magnetic Resonance Images for dogs and cats

Area of Science:

• Veterinary neurology
• Medical imaging
• Artificial intelligence in medicine

Background:

• Automated digital brain extraction is crucial for neuroimaging analysis.
• Veterinary brain extraction faces challenges due to diverse skull morphologies across species and breeds.
• Existing methods, like the atlas-based VIBE, have limitations, particularly in handling variations and have not been validated across diverse diagnoses.

Purpose of the Study:

• To develop and validate the first deep learning-based tool for robust brain extraction in dogs and cats.
• To create a tool that accounts for significant morphological variations in skull conformations.
• To outperform existing veterinary brain extraction techniques.

Main Methods:

• Development of a deep learning model trained on a cohort of 115 animals (dogs and cats).
• Validation of the tool across various skull morphologies and diagnoses.
• Performance comparison with the atlas-based VIBE method and human-centric methods.

Main Results:

• The deep learning tool achieved a Dice coefficient of 0.97 for brain extraction.
• The proposed method demonstrated robustness across diverse skull conformations and diagnoses.
• The tool significantly outperformed the previously developed VIBE method.

Conclusions:

• A novel, robust deep learning tool for canine and feline brain extraction has been established.
• This tool offers superior performance compared to existing methods, including those designed for humans.
• The validated segmentation tool has significant clinical relevance for veterinary neuroimaging applications.